MacroDiff+: Physics-Guided Diffusion for Chip Design
MacroDiff+, a physics-guided geometric diffusion framework, has been proposed for macro placement in VLSI physical design. The framework uses a dual-domain denoising architecture combining heterogeneous GNNs for topological connectivity and a Transformer for global geometric context. A Physics-Guided Sampling inference strategy ensures statistical plausibility and physical validity. On ISPD2005 MMS benchmarks, MacroDiff+ achieves a 6.1-6.2% reduction in wirelength, outperforming state-of-the-art baselines with superior stability.
Key facts
- Macro placement is a pivotal stage in VLSI physical design.
- MacroDiff+ is a physics-guided geometric diffusion framework.
- It uses a dual-domain denoising architecture with heterogeneous GNNs and a Transformer.
- Physics-Guided Sampling inference strategy is introduced.
- Tested on ISPD2005 MMS benchmarks.
- Achieves 6.1-6.2% reduction in wirelength.
- Outperforms state-of-the-art baselines.
- Exhibits superior stability.
Entities
—